Is GENERATIVE DESIGN the future of architecture?


Okay, hello everyone! I am here with Kean
and Lorenzo from Autodesk today And where are we at? …. We are at Autodesk University 2018! -That was a tricky question!
-I am sorry So maybe I’ll introduce you
you have been working for Autodesk for 23 years. – 23 years.. short years! -So what is it
exactly that you do? So for the last three years or so I’ve been in the
Autodesk research team and I focus on two main areas of research. One being the
integration of IOT data into a building model or into a 3d context
generally for visualization and the other one is on the generative design
side so I work closely with my friend Lorenzo here on that – Is it my turn now?
– Yes, please 🙂 My name is Lorenzo. I am an architect and research
scientist with The Living. The Living is a first of its kind Autodesk studio. It’s
an architectural research studio that was an independent
architecture studio that was acquired in 2014 and we’re are formerly owned by
Autodesk and part of Autodesk research My focus is on computational design and generative design.
– Very cool! So the reason I grab those two to sit down with me is because I heard a
presentation by them about generative urban design and I thought that was
really cool so I wanted to share that with you and ask them a few questions if
that is okay?
-Absolutely! – So can you briefly explain the concept of
generative urban design?
– Sure more generally speaking generative
design is a form of code design. So it involves combining on one side
artificial intelligence and on the other side human creativity to create
solutions to design problems that would otherwise not be possible. That humans
alone would have not be able to create without a computer and solutions not the
computer alone with being able to create without humans.
– Can you
maybe give an example? Sure! Another way that we describe it is instead of
starting by drawing the shape or the form of what you want you start with the
goals and the constraints of the problem and then you task the computer with the
automatic generation, evaluation and evolution of literally tens of thousands
of high-performance solutions.
– Yes, so how are the what are the steps? How does that work?
So it starts with the architect or the engineer, right?
– Yeah with any professional that is interested in changing the way we’re
currently designing. So it could be an architect, a designer or an engineer. It’s a framework first of all that can be applied to a very diverse set of
design problems but also across many different scales.
It could be applied all the way from a industrial component to architecture
so the scale buildings and all the way to the city.
– Okay and so what are
the elements that are involved with that? so there are three main steps.You start
with generation part which is a parametric model that can create a vast
solution space; so a design space of all possible solutions that that same model
can create. Then there is the evaluation part in the form of design
goals or simulation algorithms that can score numerically each design solution
and then the third component is evolution and so the optimization part
and this involves the use of a intelligent system that can learn how to
improve each design solution over generation and over generation improve their
performance
-Okay so the way that the creative process works is that you the
first a generation of tens of thousands of different designs, right? But just like
randomly?
– No actually it’s not random. It could start randomly. Are different ways
to start the first generation but soon after that the – in our case
we’re using a genetic – algorithm which is based on principles of evolution the
winning evolution. It picks the best high-performing solution of the first
generation and through this process of breeding, selection and mutation the best
performing solutions survive to the next generation and so on and so forth
you start like creating very strong high-performing solutions. Yeah so it which is totally different than random if we’re doing a totally
random search of possible solutions it would just take too long before finding
something is actually really good – And then what happens in the optimization? If
you say that the designs already evolved from generation to generation,
what happens then?
– Well what happens afterwards is the the human component
comes in in place. Well the human component is always present in the process
but specifically at the end once you have this large data set of design
options, you can start navigating the design space and through custom analysis
tools identify the trade-off between each design solution along the design
goals that you have pre-established at the beginning of the process.
– Okay well maybe Kean, you can tell me a bit more about the software that we can use in a work flow like this? Sure, sure, so on Monday this week Autodesk actually launched a public beta of Refinery which is a an extension that
you load into Dynamo which is a visual programming tool that I’m sure your
readers know about. Basically you you set up your graph inside Dynamo
with inputs and outputs marked as such inputs and outputs
and Refinery…
– What would those inputs be for example ? – Oh so just take for example I mean one
example I’ve been using is related to tiling floor plans. One of the
constraints of the system is the floor plan itself and then the inputs
could be the orientation of the tiles and then may be an offset in the x and y
distance you know directions and then so those three inputs- Varying those inputs
will of course change the outputs and in this case the outputs are the number of
complete tiles, the number of partial tiles that have to be cut to fit within
the boundary and then maybe like the discarded area of tiles that you’re
throwing away. So then based on those outputs – and Refinery will load the
inputs and the app will load the graph of the inputs and the outputs. And then you’ll
set up your your generation run and you can choose to do a completely random run
where it randomizes the various inputs you can choose an cross-product run
which is optioneering which is means that you’re sort of systematically
exploring the solutions space or the parameter space by having sort of
fixed offsets through the various parameters combined you know in with the
permutations of those and then it generates all these solutions so it’s
very systematic and then the third way of doing it is is in my opinion the most
interesting which is of course this optimization run which is using this
genetic algorithm to hone in on some optimal solutions. So typically you’ll
then specify that you want to maximize the number of complete tiles and
minimize in the discarded area of tiles and then it’ll go ahead and run
through and optimize those. It will run the graph with various inputs that
get tweaked each time or in each generation based on the outputs.
– So the example with the tiles would be something very simple that somebody
could start with, right?
– yes absolutely so I mean I’ve posted the graph for the
tiling example on my blog which I’m sure will be added to the link in the description
-Yeah!
-So go ahead and check that out -Point down and then..
-Click here for the .. no… Click here to load the dynamo
script and here to load the floorplan, does that work?
-I’ve always wanted to do that And then here you have another video … subscribe..
– Exactly
– But and so but if you say this would be a small application with the tiles, tell us how large scale can you go with this? So I mean I started I mean
even the graph for the tiles isn’t completely trivial I mean there are some steps that you
have to go through in order to do even this fairly simple problem like that. But you know the project that Lorenzo and I’ve been working on for
Van Vynen which is a different scale the urban scale. So to define
the layout for a residential neighborhood, it’s big I mean it’s a big
graph! But hey you don’t have to go big to make use of it
so I suggest that people trying it out start small and sort of think about you
know understand the system and how it works and then gradually sort of scale
up a little time.
– Well what do you.. what do you would you say to somebody who
says: Well a computer can never generate like replace the creative process of a
human being?
– So that’s.. you know I don’t think that the goal is for this
type of process to replace humans at all! I mean that’s the case but what it does
do is create insights that we as humans are sometimes kind of formatted not to
explore so it’ll explore the space in a way that we
won’t necessarily do ourselves and generate solutions that inspire us to
think about the problem differently as well So I think that that’s something
that’s very common with generative design it’s not necessarily creating a
better solution than one that a human would do but I mean it’s certainly you
know evaluated in a very numeric way but it’ll certainly give us insights that
are outside of what we would generally decide to do
because we tend to sort of work in a very structured similar way over time. So I actually find that in some ways I mean it’s not creative but it’s generating solutions that sometimes seem very creative and give
the outside of what we are used to. So that’s very interesting
and then of course you can you know over time with these systems we’ll be
factoring in machine learning and other ways which can sort of evaluate things
that are somewhat quite subjective like that’s very interesting thing about
machine learning is it’s very often they’re able to do things that seem
like they’re creative and aesthetic in nature whereas actually it’s really just
about analysis and so the combination of genetic algorithms to
search through space along with machine learning to evaluate options very
quickly I think is going to be very powerful and we’re already starting to
see some examples of those
– Like?
– Well there’s one example that I show well there’s a
video of this on the Refinery site that shows talking about optimizing the wall
to window ratio I think for the building and at the same time there’s
actually machine learning components of that which is based on some data that
was gathered through Autodesk inside 360 And so that neural network
evaluates or does a computation or you know or replaces what would be a very
complicated computation to do daylight analysis and energy analysis based on
data that you know that we require or that is grown over time so that’s one
example where you can use this combination of machine learning with
with genetic algorithms.
– You have anything you would like to add to that? – Yeah, on top of the great observations that Kean
mentioned, one important thing that we always want to stress is that this is
not a cold-blooded optimization process we are not exploiting optimization. For
us it’s more about exploration. It’s about opening it up – unlock a designer’s
creativity. So we generally also define it as a way to go beyond
intuition or rules of tongues or knowledge that you have acquired through
past experience which is very important but because of those biases that you
have as a designer you might not have thought about solving that specific
problem in a total different way. Generative design because of its nature
it’s not biased in itself as a framework Of course it is biased in
the moment there’s a designer behind it but that’s a whole different
conversation but as framework it helps you free yourself from your own biases.
– Well thank you very much for giving me those insights and sharing it with my
viewers. I feel like you’re losing your voice a little.
– I just have to cough a little bit.. but I’ll be fine
– okay so thank you very much for like even straining your voice some more – It’s really a pleasure to be here
– Thank you so much for having us
– Enjoy the rest of your stay in Vegas – You too!
– I will!
– Haha I know you will – Bye bye – Just in case: follow us here!!

Leave a Reply

Your email address will not be published. Required fields are marked *